Data-driven organizations
Best practices for AI operationalization in Sweden
Over 20 months, leading Swedish organizations from industry, academia, and the public sector joined forces to figure out how to move AI from experimentation to real-world implementation. Together, they identified the organizational, technical, and governance models that enable responsible, scalable, and efficient AI operations.
Why? Thanks to AI, organizations now have a toolbox that lets them create solutions that were previously impossible. Over the last few years, this new solution space has been explored through proof-of-concepts and pilots at many of our partners. This phase has been needed to get an understanding of how AI can create value in various real world cases.
But once you have an idea on what you want to do, you face the question on how to continue development and put models into production in a structured and organised way on scale. Hence the project Datadriven Organizations: Best practices for operationalizing AI in Sweden (DDO) was formulated.
Sustainability doesn't have to be a fancy word you mention to sound green. It's really just about not being wasteful: choosing what you actually need and optimizing what you have. The barrier to AI is lower than most of us think - what you need is smart architecture, not just computational muscle.
Milena Miernik
Developer at Aixia
Through the project, we realized that we needed to change our whole container structure. As a result of DDO, we now have a blueprint for how to scale MLops on a small number of clusters and a shared GPU pool. This will increase the utilization of investments made. We also see big value in the whitepapers produced by the other participants and will work to adapt them to our needs.
Daniel Jakobsson
Strateg Artificiell Intelligens at Trafikverket
Thanks to our participation in the project, we have received support in optimizing capacity and resources within MLOps, while also gaining valuable insights from various sectors. It has been a very positive and educational experience for the entire team.
Lina Gårdemark
Data Engineer/Data Scientist at Region Halland
Benchmark performance in a lab environment rarely translates directly to real-world use cases. Hidden bottlenecks and constraints often don't reveal themselves until you test your actual workflow – and you may find you need far fewer, or far more, resources than expected. Start by building a solid understanding of your real requirements, then match your infrastructure investment accordingly.
Olof Sandell
DevOps Engineer at Aixia
Our main learnings in the project are around how to adapt established ‘DevOps’ principles to’ MLOps’ and further on to ‘AIAppOps’, to get the whole AI ecosystem along the full model lifecycle to work together to reach a higher degree of automation – to be able to scale in number of models and thus also the value output from AI, without proportionally scaling in number of people.
Mattias Jonhede
Manager Advanced Analytics Engineering at Volvo Parts
This project has been co-funded by participating partners and Vinnova. AI Sweden is in part financed by the European Regional Development Fund under the project "Increased national collaboration and accelerated use of AI in all industries".
Case presentations
The majority of DDO’s efforts were directed towards three very concrete use cases addressing actual needs identified by project partner organizations: How to use AI in a sustainable way, how to use shared infrastructure between different applications in a way that is regulatory compliant, and how to manage and administer a situation where you have thousand of models in production?
Furthermore, more cases were explored during the project. We wish to highlight two of them in direct connection to the use-cases. One on providing guidance for organisational transformation and one on connecting the data driven aspects to an application perspective.
Sustainable AI infrastructure lifecycle
How can we make AI adoption economically viable and sustainable across its entire lifecycle, even with limited resources? Through a collaboration between Region Halland and Aixia, concrete benchmarks were conducted for both text and image analysis, comparing energy efficiency etc for hardware, models, and frameworks.
1000 models in production
How do you scale ML operations from a few to hundreds or even thousands of models while maintaining efficiency and governance without scaling the needed personnel to support the operations? Led by Volvo Parts and experts from Hopsworks, Red Hat, and Linköping University, the use case uncovered practical strategies to manage the life-cycle of countless models without proportional human resource growth. Discover strategies for successful scalable ML operations in complex business environments.
Centralized AI Infrastructure with Kubernetes: Secure and Compliant
Trafikverket needed to securely pool fragmented, specialized resources like GPUs while meeting a large array of legal requirements including MSB’s stringent segregation requirements. This use case developed a proposed robust architecture, validated through multiple proofs-of-concept utilizing Stormgrid, Red Hat and Proact technologies. Discover a secure method to share IT resources between development, test and production that enhances capacity and modernizes practices for AI-based products.
Whitepaper coming soon!
AI Apps Ops
Data-driven projects pose additional challenges to organizations due to their dependency on data across the development cycle. To aid organizations in dealing with these challenges, this whitepaper presents a framework called AI Application Operations (AI App Ops), outlining the main steps and roles involved in going from idea to production for data-driven solutions.
Building blocks for AI operationalisation
While technical capacity and pilot use-cases abound, organizations often struggle to transition from experimentation to operational deployment, not to mention taking on the challenge of becoming a fully data driven organization. The playbook consolidates strategic lessons from DDO, targeting c-level decision-makers in Sweden. It presents a leadership roadmap for embedding AI in sustainable, scalable ways – across sectors and use cases.
Whitepaper coming soon!
Whitepapers
Complementary whitepapers
A project with 20 participating organisations, 50+ individuals and on a topic as broad as MLOps, there are plenty more to dive into than the cases presented above. Below are complementary papers on specific topics that we found valuable to also explore during the project.
A Practical IT Roadmap for Enabling AI in Academia
This white paper presents a practical, experience-based roadmap that guides University IT departments in building an AI-ready environment. The roadmap emphasizes staged progress over large, speculative investments.
IBM complementary reading for DDO cases
IBM has reviewed each of the cases in the DDO project and put together a list of complementary material. The intention here is to provide more context and aid readers into making more informed decisions. Furthermore, this paper also contains a complementary review of the AI Application Operations paper on the topic of Trustworthy AI.
Optimization of Multi Agent Systems
Agents and agentic systems are peaking in interest at the moment. Not looking into agents would be a shortcoming of the project. Predli and AI Sweden has put together an introduction to multi agent systems and done a deep dive into how they can be optimized.
Technical demos, showcases and more
Complementing our white papers and the cases shared above, there is more knowledge we wish to share. In the below playlist you will find in-depth technical deep dives relevant to the cases, showcases of relevant frameworks and lessons learned from partners own MLOps journeys. These are knowledge sharing sessions from the project and provide a view directly into the dialogs and discussions had during the DDO project.
Want to explore these frameworks yourself?
As part of the project deliveries, representative setups for each vendor framework and case has been set up in the AI Sweden AI Labs. They are now available for partners for test and experimentation. In the testbed you will find Hopsworks Feature Store, IBM Fusion with Watson X, Red Hat Openshift, Red Hat Openshift AI and Stormgrids GridCloud with Run:AI.
As part of the project, IBM has installed an IBM Fusion in AI Sweden's Labs. Since it is a combination of hardware and software, below is an explanation of the components and their purpose. We have also mapped how the cases could potentially be run in a Fusion machine.
Pictures from the AI Sweden Testbed in Gothenburg. If you're part of a partner organization, please don't hesitate to contact Max Petersson, Ted Henriksson, or Laurian Lamba to learn more.
Project partners
Want to deep dive even further?
Several of the project partners provide more information relevant to DDO on their websites, explore below.
Blog post by Red Hat
The MLOps Challenge: Scaling from one model to thousands: What if managing models didn’t have to be chaotic?
Book on MLOps by Jim Dowling at Hopsworks
Building Machine Learning Systems Batch, Real-Time, and LLM Systems
An article written by Tiger et al during the project
Exploratory Visual Analysis for Increasing Data Readiness in Artificial Intelligence Projects.
AI Operation Talks: Insights from the Frontlines of Deployment
While the project defines the frameworks for governance and structure, the AI Operation Talks series showcases the practical reality of building AI-driven organizations. Curated from the vibrant Swedish AI community, this collection features "TED-style" presentations from the experts, engineers, and entrepreneurs currently navigating the shift from experimentation to production.
AI Operations talk playlist: Watch all 21 recordings from this webinar series.
Complementing the project’s whitepapers and playbooks, these talks offer deep dives into the technical and strategic hurdles of going live. Viewers can expect unvarnished lessons on MLOps, digital sovereignty, and sustainable infrastructure. From scaling operations to thousands of models to navigating the complexities of vibe coding and secure government ecosystems, these seminars provide the actionable context needed to turn strategy into operational success.
Contact information
Sign up for AI Sweden's newsletter
Newsletter subscribers will receive a notification when all whitepapers and other resources have been added to this page. You will also get monthly updates, invites, and the latest from the Swedish AI ecosystem.
Related articles
New large project to bridge the gap between AI testing and full-scale implementation
New IBM technologies create valuable test environments for AI Sweden partners
